News & Updates

Two investment analysts examine identical financial information but reach different conclusions because the data is presented in different ways. Behavioural finance has long recognised this phenomenon as framing bias. Now, research suggests that artificial intelligence can exhibit similar vulnerabilities.

As large language models (LLMs) become increasingly embedded in investment research, portfolio management and risk assessment, this raises an important question: can AI systems amplify the very cognitive biases they are expected to help investors overcome?

A new report from the CFA Institute Research and Policy Center, Managing LLM Bias in Investing: From Detection to Mitigation, by James Tait and Brian Pisaneschi, CFA, examines how biases emerge in AI-supported investment processes and evaluates practical methods for reducing their influence. 

 

Published in September 2026, the study combines behavioural finance research with an experimental analysis of how different LLMs respond to financial information, offering investment professionals a framework for building more reliable AI-enabled workflows.

When identical information produces different judgements

One of the report’s most revealing findings concerns framing bias: the tendency for decisions to change depending on whether equivalent information is presented positively or negatively.

The researchers tested nine OpenAI models across ten investment scenarios, submitting 900 prompts to each model. The scenarios covered areas including portfolio performance, drawdowns, client retention, credit defaults, concentration risk and stress testing. The objective was to determine whether changing the presentation of otherwise identical financial information would influence the models’ assessments: the results confirmed that framing matters.

The strongest and most consistent effects emerged in evaluations of drawdown performance and client retention, where positively and negatively framed information produced different assessments despite conveying equivalent underlying facts. For investment professionals, the implications are significant. An AI-generated analysis may appear objective and data-driven while remaining sensitive to the language, structure or emphasis of the information supplied. This vulnerability becomes particularly relevant when LLMs are used to summarise company results, evaluate investment opportunities or support portfolio decisions. Even when the underlying data is accurate, its presentation may influence the resulting interpretation.

Bias does not originate with the model alone

The report also challenges the tendency to treat AI bias as a problem confined to algorithms or training datasets. Instead, the authors distinguish between different sources of bias operating throughout the investment process. Human cognitive biases, including anchoring, availability and framing, can influence how professionals formulate questions, select information and interpret results.

LLMs introduce additional vulnerabilities. Implicit bias may arise from training data, model architecture and learning procedures, while explicit bias can become visible through the model’s choices about which information to prioritise, which sources to use and which analytical steps to follow. These influences can interact.

An analyst may unintentionally introduce a particular interpretation through the wording of a prompt. The model may then reinforce that interpretation by selecting supporting information or producing a persuasive explanation consistent with the initial framing. The resulting output can appear coherent and analytically sophisticated without necessarily reflecting a balanced evaluation of the available evidence.

This becomes even more important as financial institutions adopt agentic AI systems capable of performing multiple analytical steps, selecting tools and taking actions with reduced human intervention. In such environments, bias may propagate across an entire workflow rather than remaining confined to a single response.

Why telling AI to avoid bias is not enough

A particularly useful contribution of the CFA Institute research is its comparison of different mitigation techniques.

One approach might appear straightforward: explicitly instruct an LLM to recognise or avoid framing bias. The experiments, however, found that simply defining framing bias within a prompt produced little improvement. More effective results emerged when the researchers changed how information was presented.

Manually restructuring prompts to include both positive and negative perspectives produced the largest reduction in framing bias.

This human-in-the-loop approach proved more effective than relying on general instructions intended to make the model aware of its own potential biases.

The researchers also examined whether a separate LLM could identify and restructure biased prompts before they reached the model performing the investment analysis. This method successfully reduced bias in most cases and offered a potentially faster, more scalable safeguard. Nevertheless, it did not outperform direct human intervention.

The findings suggest that improving AI reliability requires more than asking models to be objective. It requires designing processes that systematically expose competing interpretations and reduce the influence of how information is initially framed.

From prompt engineering to investment governance

The practical implications extend beyond individual analysts experimenting with different prompts. As AI becomes integrated into investment operations, managing bias increasingly becomes a question of organisational governance. The report identifies several measures that financial institutions can adopt to make AI-supported analysis more transparent and reliable.

These include structured prompt templates, comparative testing across models, predefined data sources and analytical procedures, independent review mechanisms and clearly established checkpoints for human intervention.

Comparative testing is particularly important. By systematically varying the framing, context or order of information, firms can identify whether an AI system produces materially different conclusions from equivalent inputs. Internal benchmarks can then help assess the consistency of models used for recurring or high-risk tasks. For example, an investment firm using AI to support credit analysis could test whether equivalent financial indicators produce different assessments when presented through positive or negative narratives. Similarly, portfolio managers could evaluate whether changes in the order or emphasis of performance information affect AI-generated interpretations of investment results.

These controls need to be maintained over time. Model updates, changes in data sources and the introduction of new tools can alter system behaviour, making continuous testing essential. The objective is not to eliminate every possible bias, which the authors acknowledge is unrealistic, but to make biases identifiable, measurable and manageable.

Human oversight must remain meaningful

The research also reinforces the importance of human responsibility in AI-assisted investment decisions.

Human involvement is not automatically an effective safeguard. Its value depends on whether professionals have sufficient information, expertise and authority to challenge model outputs. A reviewer who simply approves an AI-generated recommendation without examining its assumptions or analytical process provides limited protection against bias.

Effective oversight requires professionals to understand how conclusions were reached, assess the evidence supporting them and recognise situations in which alternative interpretations deserve consideration. This is particularly important for consequential decisions, including portfolio allocation, credit assessments and risk management, where errors may have significant financial or regulatory implications. The report therefore recommends clear escalation procedures, documented controls and defined responsibilities for reviewing and approving AI-supported decisions.

Such measures also address a broader challenge: as investment workflows become increasingly automated, firms must ensure that accountability does not become fragmented across models, systems and human operators.

Reliability as a competitive advantage

The investment industry’s adoption of generative AI is frequently discussed in terms of productivity: faster research, more efficient data processing and the ability to analyse larger volumes of information. The CFA Institute study highlights another dimension of performance: if AI-generated conclusions can change according to how identical information is presented, the quality of an investment process cannot be measured solely by the speed or sophistication of its outputs.

Consistency, transparency and the ability to challenge analytical conclusions become equally important.

This creates an opportunity for firms that approach AI implementation as a question of investment-process design rather than simply technology adoption. Structured testing, balanced information, independent verification and meaningful human oversight can help organisations benefit from AI’s analytical capabilities without allowing hidden biases to undermine decision quality.

The central lesson is not that AI systems are uniquely vulnerable to bias. Human investors have always faced similar challenges. Rather, the research demonstrates that human and machine biases can interact, potentially reinforcing one another within increasingly automated investment processes.

For investment professionals, the challenge is therefore to build systems that recognise these vulnerabilities instead of assuming that technological sophistication guarantees objectivity. As AI becomes a permanent feature of investment management, the ability to identify and control bias may prove just as important as the ability to generate analysis itself.